The next 20 years of genome research

Michael Schatz
doi: http://dx.doi.org/10.1101/020289

The last 20 years have been a remarkable era for biology and medicine. One of the most significant achievements has been the sequencing of the first human genomes, which has laid the foundation for profound insights into human genetics, the intricacies of regulation and development, and the forces of evolution. Incredibly, as we look into the future over the next 20 years, we see the very real potential for sequencing more than one billion genomes, bringing with it even deeper insights into human genetics as well as the genetics of millions of other species on the planet. Realizing this great potential, though, will only be achieved through the integration and development of highly scalable computational and quantitative approaches can keep pace with the rapid improvements to biotechnology. In this perspective, we aim to chart out these future technologies, anticipate the major themes of research, and call out the challenges ahead. One of the largest shifts will be in the training used to prepare the class of 2035 for their highly interdisciplinary world.

Learning quantitative sequence-function relationships from high-throughput biological data

Gurinder S Atwal, Justin B Kinney
doi: http://dx.doi.org/10.1101/020172

Understanding the transcriptional regulatory code, as well as other types of information encoded within biomolecular sequences, will require learning biophysical models of sequence-function relationships from high-throughput data. Controlling and characterizing the noise in such experiments, however, is notoriously difficult. The unpredictability of such noise creates problems for standard likelihood-based methods in statistical learning, which require that the quantitative form of experimental noise be known precisely. However, when this unpredictability is properly accounted for, important theoretical aspects of statistical learning which remain hidden in standard treatments are revealed. Specifically, one finds a close relationship between the standard inference method, based on likelihood, and an alternative inference method based on mutual information. Here we review and extend this relationship. We also describe its implications for learning sequence-function relationships from real biological data. Finally, we detail an idealized experiment in which these results can be demonstrated analytically.

Optimizing error correction of RNAseq reads

Matthew D MacManes
doi: http://dx.doi.org/10.1101/020123

Motivation: The correction of sequencing errors contained in Illumina reads derived from genomic DNA is a common pre-processing step in many de novo genome assembly pipelines, and has been shown to improved the quality of resultant assemblies. In contrast, the correction of errors in transcriptome sequence data is much less common, but can potentially yield similar improvements in mapping and assembly quality. This manuscript evaluates several popular read-correction tool’s ability to correct sequence errors commonplace to transcriptome derived Illumina reads. Results: I evaluated the efficacy of correction of transcriptome derived sequencing reads using using several metrics across a variety of sequencing depths. This evaluation demonstrates a complex relationship between the quality of the correction, depth of sequencing, and hardware availability which results in variable recommendations depending on the goals of the experiment, tolerance for false positives, and depth of coverage. Overall, read error correction is an important step in read quality control, and should become a standard part of analytical pipelines. Availability: Results are non-deterministically repeatable using AMI:ami-3dae4956 (MacManes EC 2015) and the Makefile available here: https://goo.gl/oVIuE0

Mixed Models for Meta-Analysis and Sequencing

Brendan Bulik-Sullivan
doi: http://dx.doi.org/10.1101/020115

Mixed models are an effective statistical method for increasing power and avoiding confounding in genetic association studies. Existing mixed model methods have been designed for “pooled” studies where all individual-level genotype and phenotype data are simultaneously visible to a single analyst. Many studies follow a “meta-analysis” design, wherein a large number of independent cohorts share only summary statistics with a central meta-analysis group, and no one person can view individual-level data for more than a small fraction of the total sample. When using linear regression for GWAS, there is no difference in power between pooled studies and meta-analyses \cite{lin2010meta}; however, we show that when using mixed models, standard meta-analysis is much less powerful than mixed model association on a pooled study of equal size. We describe a method that allows meta-analyses to capture almost all of the power available to mixed model association on a pooled study without sharing individual-level genotype data. The added computational cost and analytical complexity of this method is minimal, but the increase in power can be large: based on the predictive performance of polygenic scoring reported in \cite{wood2014defining} and \cite{locke2015genetic}, we estimate that the next height and BMI studies could see increases in effective sample size of $\approx$15\% and $\approx$8\%, respectively. Last, we describe how a related technique can be used to increase power in sequencing, targeted sequencing and exome array studies. Note that these techniques are presently only applicable to randomly ascertained studies and will sometimes result in loss of power in ascertained case/control studies. We are developing similar methods for case/control studies, but this is more complicated.